Easy AI Daily | January 28, 2026

Key points

• Kimi K2.5 released: Moonshot's open-source multimodal MoE (32B active / 1T params) tops open-model leaderboards on HLE, BrowseComp, MMMU Pro, VideoMMMU, and SWE-bench. Native image+video understanding, can reconstruct webpages from screen recordings, 128K→256K context, INT4 partial quantization; available on HuggingFace, Ollama, Together, Fireworks, and runnable locally on multi-GPU Macs. Tech blog
• Trinity Large preview: Arcee + Prime Intellect + Datology release a 400B MoE (13B active) trained on 17T tokens — 3:1 local/global gated attention, SWA, NoPE+RoPE, Muon optimizer, ~2000 B300 GPUs for a month. vLLM day-one support; free on OpenRouter. Announcement
• DeepSeek-OCR 2 open-sourced: Introduces Visual Causal Flow and DeepEncoder V2, compressing images to ~256–1120 vision tokens; scores 91.09% (+3.73) on OmniDocBench v1.5. vLLM supported. Model
• OpenAI Prism: Free GPT-5.2-powered research workspace with LaTeX writing, collaboration, citation, and literature search; no automatic IP claims on research output. Launch
• Qwen updates: Qwen3-Max-Thinking claims near-commercial closed-model reasoning (community reports code agent compilation issues); Qwen3-32B INT4 on H100 loses only ~1.9% on MMLU-Pro while concurrent users rise from 4 to 47.
• Transformers v5: 6–11x MoE prompt speedups, faster dynamic weight loading, unified tokenizer/backend config; community tests show ~50% faster single-request inference and 2x concurrent throughput. Repo

Agents and tooling

• Kimi Agent Swarm (beta for paid users): dynamically spawns up to 100 sub-agents, up to 1500 tool calls, end-to-end task time reportedly cut to 1/4–1/3 of a single agent; trained with parallel agent RL (PARL).
• Kimi Code & Agent SDK: Apache-2.0 open-source coding agent with IDE integration plus an SDK for custom agent workflows.
• Converging multi-agent patterns: Kimi Swarm, LangChain sub-agents, and Google Jules' Planning Critic share an architecture of orchestrator + parallel sub-agents + plan reviewer; Jules claims ~9.5% lower task failure rate.
• IDE agents: Cursor reports big quality/speed gains from semantic search indexing of large codebases; VS Code adds command-execution explanations and MCP Apps returning UI components.
• Community practice: A "hive" of 7 Claude Code agents sharing SQLite+FTS5 memory via MCP; comparisons with Microsoft Autogen and BMAD; CheshireCat Enterprise targets multi-tenant agent workflows.
• Karpathy goes agent-first: Andrej Karpathy says he's shifting workflows to LLM agents (Claude etc.) that continuously code and run experiments, with humans supervising — "models never get tired."

Infrastructure and hardware

• Unsloth: MoE training now ~14x faster than v4 via new kernels + Transformers v5, targeting 30x; focused on consumer/cloud GPU training of large MoEs.
• FlagOS (GPU MODE community): open system software stack aiming to unify model–system–chip layers for portable, tunable AI workloads.
• tinygrad / megakernels: automatic rewriting of naive attention into FlashAttention kernels; exploration of compiling models into single megakernels; George Hotz cites Luminal's vision of GPUs running OS-like kernels. Blog
• Multi-GPU budget LLM rigs: Used Tesla GPUs at 200GB+ VRAM — bottlenecks are PCIe bandwidth and prefill speed, plus thermal/power costs; few benchmarks cover sliced multi-GPU serving.
• FlashInfer-Bench: Inference trace and MLSys 2026 contest workload datasets released on HuggingFace, with biweekly leaderboards planned. Traces

Research and methods

• Hallucination inevitability: A paper argues current-paradigm LLMs can never fully eliminate hallucinations; jailbreaks exploit this by shifting context so the model no longer flags unreliable content. arXiv:2409.05746
• Anthropic bio-risk finding: Light fine-tuning of open models on frontier closed-model outputs restores or enhances suppressed dangerous capabilities at low compute cost — "refusal-only" safety is fragile. PDF
• Planning & RL: DeepPlanning benchmark shows agents still struggle at verifiable multi-day planning; PrefixRL reuses old trajectories via prefix conditioning for ~2x faster RL convergence.
• Scaling & math: Google's ATLAS gives multilingual LLM scaling-law guidance on data mix and model size; Epoch's FrontierMath: Open Problems — no AI has yet solved any included genuine research-level math problems.
• MergeMix: Learnable model merging to auto-search data mix ratios mid-training; useful for compute-constrained open projects. Paper

Products and applications

• Kimi K2.5 office agent: Popular for long-document report writing and summarization in China; some companies use it to replace weekly-report/analysis workflows.
• Gemini AI Studio cutbacks: Free tier reduced; Pro/Ultra reportedly have only 32k–128k effective "hot memory" vs advertised million-token context, pushing users to Grok 4.1 (2M) or Claude Sonnet 4.5 (1M), or toward retrieval-augmented setups.
• Perplexity Pro friction: Rate-limited search/image generation and opaque billing (especially payment failures in India) push some users toward privacy-focused Kagi.
• Local coding assistants: On 8GB VRAM + 32GB RAM, Qwen2.5/3 Coder 7B/30B are the practical ceiling; multi-agent plugins like Cline frequently hit CUDA OOM errors without careful tuning.

Industry and governance

• Chinese open models vs Western closed: Kimi K2.5 approaches or beats Claude Opus 4.5 / Gemini 3 Pro on multiple agentic, coding, and multimodal benchmarks as open weights; a16z reports 80% of startups use Chinese open-source models.
• Pre-holiday release wave: MiniMax teases M2.2; Qwen hints at new vision/multimodal models (Z-Image) via ComfyUI PRs.
• Clawdbot → Moltbot: Renamed over Claude trademark conflict; worse, community reports it can read environment variables without authorization — several communities now advise against it. Video
• Decart Lucy 2: Autoregressive video editing model released; hiring kernel engineers for real-time video/world models on accelerators like Trainium 3.
• AI detector false positives: Detectors flag pre-ChatGPT papers as AI-generated, yet schools and recruiters still rely on them.
• Gemini "consumer fraud" claims: Experiments suggest ~32k real hot context for Gemini Pro, plus billing bugs (one user reportedly charged over $70k by mistake).
• GPT-5 "control shell" leak: A file (GPT5_Hotfix.md) posted in BASI Jailbreaking claims to be a pre-generation control layer with syntax constraints, intent locking, and drift protection — if genuine, frontier labs increasingly rely on external rule layers.

📌 Source: Easy AI Daily (zhichai.net)